Keras convolution shape has dimensions out of order (Error when checking model input)

Question

I've checked all the similar posts, but my error isn't being fixed with the fixes suggested. Thanks in advance for any help!

I'm using a tensorflow backend with Keras, and my images have dimensions 1185 by 676. Most of the code is from one of the Keras examples.

I'm getting ValueError: Negative dimension size caused by subtracting 2 from 1 for 'MaxPool' (op: 'MaxPool') with input shapes: [?,1,1183,32]. This error disappears when I switch to dim_ordering="th", which is odd, considering that I'm using tensorflow, not theano.

The code up to this point:

img_width, img_height = 1185, 676

train_data_dir = 'data/train'
validation_data_dir = 'data/validation'
nb_train_samples = 32
nb_validation_samples = 8
nb_epoch = 3

model = Sequential()
model.add(Convolution2D(32, 3, 3, input_shape=(3, img_width, img_height)))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2), dim_ordering="tf"))

And just in case the data generation is part of the issue:

model.compile(loss='binary_crossentropy',
              optimizer='rmsprop',
              metrics=['accuracy'])

train_datagen = ImageDataGenerator(
        rescale=1./255,
        shear_range=0.2,
        zoom_range=0.2,
        horizontal_flip=True)

test_datagen = ImageDataGenerator(rescale=1./255)

train_generator = train_datagen.flow_from_directory(
        train_data_dir,
        batch_size=4,
        target_size=(img_width, img_height),
        class_mode='binary')

validation_generator = test_datagen.flow_from_directory(
        validation_data_dir,
        batch_size=4,
        target_size=(img_width, img_height),
        class_mode='binary')

model.fit_generator(
        train_generator,
        samples_per_epoch=nb_train_samples,
        nb_epoch=nb_epoch,
        validation_data=validation_generator,
        nb_val_samples=nb_validation_samples)

Update: It looks like the dimensions weren't divisible by the pool size. Now I'm getting `ValueError: Negative dimension size caused by subtracting 3 from 1 for 'Conv2D' (op: 'Conv2D') with input shapes: [?,1,1185,676], [3,3,676,32].` — daniel.rigberg, Jan 18 '17 at 00:04
Update: Fixed the pool sizes. Now, during fitting: `ValueError: Error when checking model input: expected convolution2d_input_1 to have shape (None, 3, 1185, 676) but got array with shape (2, 1185, 676, 3)` — daniel.rigberg, Jan 18 '17 at 00:10
Your input shape is inconsistent with image_ordering = "tf", why are you forcing an image ordering in just one layer? — Dr. Snoopy, Jan 18 '17 at 09:01
Ah!!! I can't believe that I didn't see that sooner--that's exactly the problem. Thank you! — daniel.rigberg, Jan 18 '17 at 15:17

score 2 · Accepted Answer · edited May 23 '17 at 12:16

2

The image dimension ordering has been mixed in your code. Multiple way to fix this.

One way is to add

from keras import backend as K
K.set_image_dim_ordering('tf')

at the beginning of your code.

Other approaches are summarized in this answer

edited May 23 '17 at 12:16

Community

1
1

answered Jan 18 '17 at 15:17

pyan

3,577
4
23
36

1

Thank you! I'd vote for your response, but I don't have enough reputation yet for voting. <3 – daniel.rigberg Jan 29 '17 at 21:10
1

No problem, as long as it helped. – pyan Jan 30 '17 at 15:56

Keras convolution shape has dimensions out of order (Error when checking model input)

1 Answers1